Implementation of deep learning in low-cost hardware, such as an edge device, is challenging. Reducing the complexity of the network is one of the solutions to reduce resource usage in the system, which is needed by low-cost system implementation. In this study, we use the general average pooling layer to replace the fully connected layers on the convolutional neural network (CNN) model, used in the previous study, to reduce the number of network properties without decreasing the model performance in developing image classification for image search tasks. We apply the cosine similarity to measure the characteristic similarity between the feature vector of image input and extracting feature vectors from testing images in the database. The result of the cosine similarity calculation will show the image as the result of the searching image task. In the implementation, we use Raspberry Pi 3 as a low-cost hardware and CIFAR-10 dataset for training and testing images. Base on the development and implementation, the accuracy of the model is 68%, and the system generates the result of the image search base on the characteristic similarity of the images.
A novel approach to enable semantic and visual image summarization for exploratory image search
